nekagit/ai-cli • Core Systems

MiniLLM Pusher: Adaptive Local Model Benchmark

Autonomous hardware sensing, task granularity balancing, and high-throughput Mini-LLM pushing across Apple Silicon (M4) and AMD Strix Halo (128GB LPDDR5X).

🚀 Live Deployments on nalan-server (195.90.217.100)

Both generated legal sites were autonomously provisioned with TLS certificates via ai host:

Lawyer 1 (MacBook Air Constrained Build) https://lawyer1.karaweiss.org ↗

10 fine-grained simple tickets (Difficulty 1..3) • Executed locally on Apple M4.

Lawyer 2 (KW Nucbox Capable Build) https://lawyer2.karaweiss.org ↗

4 consolidated feature capsules (Difficulty 4..5) • Executed via ROCm Qwen3-Coder-Next (:8086).

Experiment Hub (This Dashboard) https://minillmpusher.karaweiss.org ↗

Hosted static document root /var/www/minillmpusher with TLS.

Core Philosophy
Mini-LLM Pushing

Decompose any product intent into high-volume, low-cognitive-load tickets (Difficulty 1–3) to finish 70–80% of development locally on <5B models.

Hardware Range Tested
16GB → 128GB

Tested on Apple M4 MacBook Air (16GB Unified) and GMKtec NucBox EVO-X2 (AMD Ryzen AI Max+ 395, 32 threads, 121GB RAM, Radeon 8060S ROCm).

Pass Rate & Verification
100% Pass

Statutory German telemedia compliance (§ 5 DDG, DL-InfoV, BRAO), DSGVO privacy, and Schema.org LegalService structured data.

🖥️ Node Hardware Profiles

Parameter Node 1: MacBook Air M4 Node 2: GMKtec NucBox EVO-X2 Hosting Node: Nalan Server
Host Identifier Mac.fritz.box kw-NucBox-EVO-X2 (Tailscale) nalan-server (195.90.217.100)
Architecture & OS darwin / arm64 (macOS) linux / amd64 (Ubuntu 24.04) linux / amd64 (Ubuntu 22.04 LTS)
Compute Cores 10 cores (4P + 6E) 32 threads (AMD Strix Halo) AMD EPYC / KVM Virtualized
System Memory 16 GB Unified (~8 GB available) 121 GB LPDDR5X (~68 GB free) 8 GB RAM
Acceleration Apple Metal AMD ROCm / HIP (Radeon 8060S) CPU Only
Detected LLM Endpoints Mini-model direct runner :8086 (Qwen3-Coder-Next 80B MoE) N/A (Web Edge Gateway)
Assigned Device Tier CONSTRAINED CAPABLE EDGE GATEWAY

🧪 Experiment Test Rounds

Round 1: Capacity Balancing Test COMPLETE

Demonstrated adaptive capacity balancing between constrained and workstation nodes:

  • MacBook Air (Constrained): 10 fine-grained simple tickets (d in [1, 3]) mapped to qwen2.5-coder-1.5b.
  • KW Nucbox (Capable): 4 consolidated feature capsules (d in [4, 5]) mapped to qwen3-coder-next (80B MoE on :8086).
# Output manifests:
round1-mac-tickets.md    (10 tickets, diff 1..3)
round1-nucbox-tickets.md (4 tickets, diff 4..5)
Round 2: Universal Mini-LLM Pushing COMPLETE

Configured Mini-LLM Pushing Mode as universal default. Both hosts receive the identical fine-grained ticket set:

  • MacBook Air: 10 fine-grained tickets (all d in [1, 3]).
  • KW Nucbox: Identical 10 fine-grained tickets (all d in [1, 3]).
  • Obsidian Integration: Seeded KWDEV/AI-CLI-Setup.md note into Obsidian vaults on both nodes.
# Output manifests:
round2-mac-tickets.md    (10 tickets, diff 1..3)
round2-nucbox-tickets.md (10 tickets, diff 1..3)

📋 Standard 10-Ticket Mini-LLM Decomposition

For the Dortmund lawyer task, the system generated and passed all 10 modular tasks:

ID Title Difficulty Provider Model Target Outcome
LAW-0001 HTML5 Semantic Scaffold & Header 1/10 Local qwen2.5-coder-1.5b Passed
LAW-0002 Hero & Kanzlei Profile Section 2/10 Local qwen2.5-coder-1.5b Passed
LAW-0003 Legal Practice Areas (Rechtsgebiete) 2/10 Local qwen2.5-coder-1.5b Passed
LAW-0004 German Legal Impressum (§ 5 DDG / § 2 DL-InfoV) 2/10 Local qwen2.5-coder-1.5b Passed
LAW-0005 DSGVO Privacy Notice (Datenschutzerklärung) 2/10 Local qwen2.5-coder-1.5b Passed
LAW-0006 Appointment Request Form 2/10 Local qwen2.5-coder-1.5b Passed
LAW-0007 Dortmund Geo-SEO & Schema.org LegalService 3/10 Local qwen2.5-coder-1.5b Passed
LAW-0008 Responsive Law Firm CSS Design System 2/10 Local qwen2.5-coder-1.5b Passed
LAW-0009 Client-side Validation & Notification Logic 2/10 Local qwen2.5-coder-1.5b Passed
LAW-0010 Test Suite & Regulatory Compliance Verification 2/10 Local qwen2.5-coder-1.5b Passed

🔍 Automated Verification Log

=== Kanzlei am Friedensplatz Dortmund: Legal & Technical Verification ===
1. Checking Core Web Assets:
  [PASS] File exists and non-empty: index.html
  [PASS] File exists and non-empty: impressum.html
  [PASS] File exists and non-empty: datenschutz.html
  [PASS] File exists and non-empty: styles.css
  [PASS] File exists and non-empty: app.js
  [PASS] File exists and non-empty: TICKETS.md
2. Checking Dortmund Legal & Regulatory Requirements:
  [PASS] Telemedia regulatory statutory reference (§ 5 DDG) in impressum.html
  [PASS] Supervisory Bar Association (RAK Hamm) in impressum.html
  [PASS] Federal Lawyers' Act (BRAO) in impressum.html
  [PASS] General Data Protection Regulation (DSGVO) in datenschutz.html
  [PASS] Dortmund locality and court jurisdiction in index.html
  [PASS] Schema.org structured data in index.html
  [PASS] Schema.org LegalService entity in index.html
3. Result:
>>> ALL VERIFICATION CHECKS PASSED SUCCESSFULLY. Kanzlei Dortmund site is compliant.